license: apache-2.0
base_model: huihui-ai/Huihui-Qwen3.8-27B-abliterated
tags:

• georefine
• lossless-compression
• qwen3_5

Qwen3.8-27B GeoRefine Context — abliterated parent

The pinned Huihui abliterated model is stored with GeoRefine GLC v2 reversible
lossless compression. Decoding recovers every parent BF16 tensor bit for bit,
including the retained text, vision and MTP components.

Parent: huihui-ai/Huihui-Qwen3.8-27B-abliterated at
739e3c5b89849f6c238ce1e5b70008612ae42cdd.

The complete pinned parent contains 55,586,059,962 file bytes. This release
contains 36320115845 file bytes, including model frames,
metadata, decoder, licenses and release evidence. Its complete-package
compression ratio is 1.530448×, verified against an exact
minimum of 1.53×. Checksums and source provenance are included.

Verify and restore

The Linux reference decoder uses Python, NumPy and a C++17 compiler. After
downloading the complete repository at the release revision:

python package/decoder/predictor_package.py verify package
python package/decoder/predictor_package.py restore package /data/new-parent-bf16

Restoration requires space for the full BF16 parent. The decoder verifies each
tensor and the reconstructed safetensors shard hashes. Use a new destination.
The restored files can be loaded through the parent's supported Transformers
architecture and its original chat template.

This release precedes the separate dynamically quantised phone model. Phone
memory, latency and accelerated inference measurements for this parent remain
pending. The publisher describes the parent as abliterated; no universal
zero-refusal claim has been established by these compression checks.

The parent license and metadata are retained in model-assets.tar.xz within
package/. GeoRefine codec source and release provenance accompany this model.